Skip to main content

TWIN lift ball: closed-chain bimanual motion

This example adapts the TWIN/PerAct2 “lift ball” task to two Franka Panda arms. Each arm grasps a distinct handle on the same ball; the pair then raises it while both rigid grasp constraints remain active.

Fresh TWIN initial-scene capture with two Panda arms and the shared ball

Fresh viser capture from the loaded 25-DOF scene, 27 September 2026.

Planning problem​

The model has two independent seven-axis arms with articulated two-joint grippers and one free-flying ball. Finger joints remain fixed at a useful rendered width because grasp closure is represented by a rigid TCP constraint.

Why moving the ball directly fails​

After both grasps activate, the ball pose is dependent on both arm configurations. Adding 10 cm to the ball’s free-flyer Z value creates an inconsistent guess; projection can return the ball to the same constrained pose with zero motion.

The implementation perturbs both arms’ panda_joint2, projects that guess onto the active dual-grasp node, and keeps the direction that raises the ball. It repeats in small steps because a single large projection may not converge and because the useful joint sign depends on the sampled elbow configuration.

for sign in (1, -1):
    q_guess[left_joint] += sign * step
    q_guess[right_joint] += sign * step
    ok, q_projected = config_gen.project_on_node(node, q_guess)
    if ok and ball_z(q_projected) > ball_z(q_current):
        keep(q_projected)

The resulting configuration already satisfies both grasps. plan_loop then searches for a collision-free path inside that constrained state.

Execution​

python script/twin/task_lift_ball.py --backend pyhpp
python script/twin/task_lift_ball.py --backend pyhpp --viewer

Planning completes before the viser server starts. This ordering avoids a reproduced native concurrency failure when WebSocket visualization and the RRT/collision checker ran simultaneously.

Latest observed run​

The 27 September 2026 run loaded the scene successfully and planned the left-arm grasp. Its first phase took 80.92 s. The second-arm pregrasp target projected, but all path attempts failed because panda_left/panda_leftfinger_2 collided with ball/base_link_0. The process exited before dual grasp and lift.

This result narrows the current issue to collision semantics around the already-held left grasp. It also means the former wording that the example demonstrated a completed cooperative lift was too strong. The implementation supports that intended flow, while the current checked configuration did not complete it.

Limitations and next measurement​

The implementation reports the two grasp phases independently from the final constrained lift, so a failed lift does not erase earlier phase evidence. The source contains no pinned multi-seed result set or completed replay media. Accordingly, this example currently demonstrates scene loading, phase-graph construction, and one-arm grasp planning; it is not yet a successful bimanual benchmark.

A complete evaluation should record grasp success, achieved vertical displacement, closed-chain planning time, path length, minimum collision clearance, and failure classification across fixed seeds.